structure-prediction

name	structure-prediction
description	Protein structure prediction from sequence. ESMFold-based, single GPU, no MSA needed. Predicts 3D structures with pLDDT confidence scores for drug discovery targets.
category	chemistry
license	MIT
metadata	{"skill-author":"Synthetic Sciences"}
version	1.0.0
author	Synthetic Sciences
tags	["Protein Structure","ESMFold","Bioinformatics","Deep Learning"]
dependencies	["fair-esm","torch>=1.12.0","biopython>=1.84"]

Structure Prediction (ESMFold)

Overview

This skill provides protein 3D structure prediction from amino acid sequences using ESMFold (Evolutionary Scale Modeling Fold). ESMFold is a single-sequence protein structure prediction model developed by Meta AI that produces accurate 3D coordinates directly from an amino acid sequence without requiring multiple sequence alignments (MSA) or template search.

Key advantages of ESMFold:

• No MSA required: Predictions run on a single sequence, making inference dramatically faster than AlphaFold2 (seconds vs. minutes/hours).
• Single GPU execution: The entire model fits on one GPU (requires ~16 GB VRAM for sequences up to ~400 residues, more for longer sequences).
• End-to-end: Takes a raw amino acid string and outputs a full PDB structure with per-residue confidence scores (pLDDT).
• Drug discovery ready: Suitable for rapid screening of target structures, variant modeling, and initial structural hypotheses.

When to Use This Skill

Use the structure-prediction skill when you need to:

• Predict a protein structure from an amino acid sequence
• Fold a protein when no experimental structure is available
• Screen multiple sequences for structural viability in batch mode
• Evaluate prediction confidence to assess reliability of modeled regions
• Compare a predicted structure against an experimental reference (e.g., from PDB)
• Identify disordered regions in a protein based on low-confidence scores
• Generate initial models for downstream molecular docking or dynamics simulations

Trigger phrases: "predict structure", "fold protein", "run ESMFold", "structure from sequence", "batch fold", "evaluate pLDDT", "compare structures"

Related Skills

• alphafold-database: If the protein has a UniProt ID, check AlphaFold DB first — pre-computed structures are instant, no GPU needed.
• esm: For generative protein design, embeddings, or inverse folding (broader than just structure prediction).
• protein-diagram: For rendering 2D diagrams of protein structures (domain maps, Ramachandran plots) — not prediction.

Installation

pip install fair-esm torch biopython

Requirements: